China's H1 2026 AI Scorecard: Why Qwen's 5,793% Growth Is a Trap for GEO Budgets

Qwen grew its monthly active users by 5,792.9% year-over-year in June 2026 — the fastest growth of any AI-native app in China. In the same report, its average session duration per user fell 14.1%. Those two numbers appear in the same QuestMobile dataset, and only one of them makes it into vendor decks.

For brands allocating H2 2026 budget across China's AI models, this gap is the whole story. MAU tells you how many people opened an app. It does not tell you how many questions they asked, and questions — not installs — are what generate the brand mentions you are trying to win. When a platform doubles its user base while its per-user engagement contracts, the number of shots at being cited grows far more slowly than the headline suggests.

Here is what the H1 2026 data actually supports.

The June 2026 scoreboard

QuestMobile's June ranking of China's AI-native apps puts total category MAU at 499 million, up 85.4% year-over-year. The distribution underneath that number is extreme:

Doubao is the only app in the hundred-million-plus tier that QuestMobile describes as "断层第一" — first by a discontinuity, not a margin. Qwen is second but sits at 43.7% of Doubao's user base. DeepSeek, the app that defined China's 2025 AI narrative internationally, is now third and shrinking. Kimi has fallen to ninth with under 20 million MAU and the steepest decline in the top ten.

If you built a China GEO plan in Q1 2026 and weighted it by the rankings as they stood then, it is now wrong in at least two places.

The number the headlines skip: interaction volume

QuestMobile also reports per-user intensity. Doubao users averaged 76.7 sessions and 143.7 minutes per month in June.

Multiply that out and Doubao alone accounts for roughly 29.3 billion sessions per month — 382 million users × 76.7 sessions. That is the denominator that matters for brand visibility. Every one of those sessions is a chance for a model to name a brand, cite a source, or surface a product card.

Now apply the same logic to Qwen. Its MAU multiplied by nearly 59× year-over-year, but QuestMobile reports session count up only 5.2% and duration down 14.1% on a per-user basis. A user base that grows 59× while per-user depth shrinks is not a user base that has arrived — it is one that has been acquired. The interaction volume Qwen generates is real and rising, but it is rising on the back of shallow sessions from users who were subsidised or funnelled in, not from users who have made it their default place to ask questions.

DeepSeek and Kimi are the mirror image: QuestMobile shows both with negative year-over-year growth on session count and duration alongside their MAU declines. That is a compounding loss. Fewer users, each asking fewer questions.

There is a third figure worth extracting from the same table. Add the individual app MAUs together — 382M plus 167M plus 130M plus 49.8M plus 29.0M, before the sub-20M tail — and you get roughly 758 million against a deduplicated category total of 499 million. The overlap is about 260 million users, which means the average Chinese AI user is running at least one and a half of these apps, and the heavy users are running three or four.

For brands, multi-homing is the argument against single-platform optimisation. The same consumer researching a purchase will plausibly ask Doubao, then cross-check with Qwen or DeepSeek before buying. Winning one platform and losing the other two does not get you a third of the outcome — it gets you a contradiction the shopper notices. Consistency of how your brand is described across models is a separate objective from raw score on any one of them, and it is the objective most brands are not measuring.

Why Qwen's curve is subsidy-shaped

A 5,792.9% increase does not happen organically. Alibaba spent H1 2026 pushing Qwen through every distribution surface it controls — the Spring Festival subsidy campaign, Quark integration, and the Taobao commerce loop that went live ahead of 618. The company bought the entrance; it has not yet demonstrated it can hold the habit.

This distinction has a direct budget consequence. Brands that read "fastest-growing AI app in China" and reweighted toward Qwen content in Q2 are paying for reach into sessions that are getting shorter. Brands that instead asked "where are people asking the most questions in my category?" landed somewhere different.

The honest position for H2 2026: Qwen is a genuine second platform and worth optimising for, particularly in commerce-adjacent categories where the Taobao integration gives it structural advantage. But it is not a co-equal of Doubao, and the growth rate is not evidence that it will become one. Watch whether per-user duration stabilises in Q3 — that is the leading indicator of whether the acquisition converted.

DeepSeek and Kimi are declining for different reasons

Treating both as "losing" is a modelling error.

DeepSeek's 20.3% MAU decline is a consumer-app decline. The model itself remains widely deployed — through APIs, through WeChat, through enterprise integrations that do not appear in an app-install ranking. A brand that writes DeepSeek off because its standalone app is shrinking will lose visibility in every downstream surface that runs DeepSeek underneath. The app is a poor proxy for the model's reach.

Kimi's 42.7% decline is closer to a genuine category exit at the consumer level. Moonshot's long-context and agentic-coding positioning has real traction with developers, but developer traction does not produce consumer brand-recommendation queries. If your category is consumer goods, Kimi's weight in your GEO plan should be small and shrinking. If you sell developer tools or enterprise software, the calculation inverts.

The commerce layer is changing what "being cited" means

Two things happened in Q2 2026 that alter the value of a citation.

Doubao launched paid tiers — a professional subscription priced up to RMB 500 per month, activated in late June alongside ByteDance's Force developer conference. ByteDance has signalled deeper Douyin e-commerce integration in Q3, extending the connection first built in October 2025 that lets Doubao embed horizontal product recommendation cards inside conversational answers.

Qwen has the parallel arrangement with Taobao.

This means the surface you are optimising for is bifurcating. A text mention in an informational answer and a product card in a commerce-intent answer are different placements with different mechanics, and the second one is likely to become partly purchasable. Paid ranking inside AI answers is the next fight in China's AI commerce, and brands that treat GEO purely as an organic-content discipline will find the highest-intent queries fenced off behind a placement auction they did not budget for.

The practical read: earned citation in informational queries remains the durable asset, because it is the layer least likely to be monetised away. Commerce-intent queries should be planned as a paid channel with an organic floor, not as pure GEO.

Takeaway for Brand Marketers

Stop budgeting on MAU. Budget on estimated query volume in your category. MAU × session frequency is a crude but far better proxy. On that basis, Doubao's lead over Qwen is materially wider than the 382M-vs-167M ranking implies, because Doubao's engagement is growing while Qwen's per-user depth is falling.

Treat Qwen as a second platform, not a co-leader — and re-evaluate in Q3. Optimise for it, especially if you sell through Taobao, but do not reweight your plan on the 5,793% figure. The number that will tell you whether Qwen has arrived is per-user session duration, not MAU.

Do not confuse DeepSeek's app decline with DeepSeek's model reach. Keep DeepSeek-facing content investment intact if your buyers touch WeChat, enterprise tools, or any product with DeepSeek underneath. Audit where the model appears, not where the app ranks.

Reduce Kimi weight for consumer categories, hold it for B2B. Under 20M MAU and −42.7% is not a consumer discovery surface any more. It remains relevant for developer and technical-buyer audiences.

Split your query set by intent before you split your budget. Informational queries are the earned-media layer and should absorb the bulk of your content investment. Commerce-intent queries inside Doubao-Douyin and Qwen-Taobao are drifting toward paid placement — plan them with your media team, not your content team.

Re-baseline quarterly. In six months, DeepSeek went from category-defining to third and falling, Kimi went from top-tier to ninth, and Qwen went from negligible to second. A GEO plan built on the Q1 2026 ranking is already stale. The platform mix is moving faster than most annual planning cycles.

Related: See how brands in your category currently score across China's six major AI models on the hubGEO brand index.

Sources: QuestMobile June 2026 AI-native app MAU ranking; QuestMobile 2026 H1 AI application market report (via 36Kr); Sina Finance; BigGo Finance and China Biz Insider on Doubao monetization.